Last Friday, Australian data center operator Firmus, backed jointly by Nvidia and Blackstone, announced it was withdrawing its listing plan, which would have been Australia's second-largest IPO in history. The company had sought a valuation of up to A$44 billion (approximately US$30.5 billion)—exactly double its valuation from a funding round three months earlier, and far exceeding the market capitalization of established peer NextDC, which has nearly A$500 million in revenue and A$82 million in net profit. Days later, another Australian data center company, AirTrunk, also canceled its plan to list a REIT vehicle in Singapore, citing "market conditions."
The two failed listings quickly sent ripples through global markets. CoreWeave CEO Mike Intrator, in an internal memo on Thursday, attributed part of the company's 17% stock decline over the week to these events: "The failed IPOs and debt financing difficulties of companies like Firmus, AirTrunk, Lambda Labs, and Nscale are bleeding into the market narrative surrounding CoreWeave." He emphasized that these companies were trying to raise massive amounts of capital without "execution, revenue, and scale," which he said was entirely different from CoreWeave's situation.
However, this defense did not ease market concerns—CoreWeave bonds still yielded close to 15% on Thursday, with bond prices falling below the 80-cent distressed threshold.
Anatomy of Firmus's Failure: Greed, Collapsed Trust, and a Valuation Bubble
Firmus's failed IPO is a textbook case of runaway valuations and repeatedly damaged trust.
The company started in Tasmania in 2019 as a Bitcoin miner, later pivoted to AI infrastructure, and gained Nvidia's backing. Co-founder Oliver Curtis's background—having been convicted and imprisoned for insider trading in 2016—made some potential shareholders uneasy. More critically, the US$30 billion valuation Firmus sought was double its valuation from a funding round three months earlier, and the up to US$30 billion in debt it planned to take on after listing would push its debt/EBITDA ratio to six times its projected 2028 earnings. Some superannuation funds, such as UniSuper, refused to participate outright, citing the valuation as too high.
The final straw that broke the IPO was the revelation, one week before listing, that the "Project Southgate" deal had fallen apart—a multibillion-dollar cooperation agreement between Firmus and CDC suddenly collapsed, directly shaking potential investors' confidence in the company's ability to execute commercially.
Jun Bei Liu, founder and chief portfolio manager of investment firm Ten Cap, said bluntly at Morgan's annual investment conference: "They were too greedy and should have been more realistic." Underwriters Morgan Stanley, JPMorgan, and Bank of America all declined to comment publicly on the withdrawal.
Firmus now says it will turn to private market financing, but in the current credit environment, its borrowing costs may rise sharply. Meanwhile, SharonAI, which is also listed on Nasdaq and seeking a secondary listing in Australia, has seen its share price fall nearly 20% over the past five days—the chill in Australia's data center capital market is far from dissipating.
Bond Market: Pressure Spreads from the Periphery to the Core
Firmus's failure was merely the fuse; the real powder keg had long been laid.
According to an October 16 report by the UK's Financial Times, bond prices for well-known data center projects providing computing power to tech giants such as Meta, Nvidia, and Oracle have recently fallen across the board below par: the "Hyperion" project bond backed jointly by Blue Owl and Meta fell from above par in early July to about 92 cents; Blackstone's Georgia QTS "Project Excalibur" bond, leased to Microsoft, has fallen below 88 cents; Hut 8's Louisiana and Texas project bonds fell to about 94 cents and 92 cents, respectively. Oracle-related projects have come under especially severe pressure—about US$18 billion in loans tied to the "Project Jupiter" New Mexico data center have slid into distressed territory due to power permit delays, and the Michigan AI campus construction bond has also fallen from par to 87 cents.
Neocloud companies have been hit even harder. CoreWeave bonds have fallen below the 80-cent distressed threshold, with yields approaching 15%, while competitor Nebius fell 18% this week. Nvidia has shed about US$270 billion in market value over the past week.
Ludovic Subran, chief investment officer at Germany's Allianz, characterized the situation as a "Darwinian moment": "The surge in credit spreads triggered by power and construction delays shows that data center debt pricing has been too tolerant of execution risk."
The US$4.9 Trillion Gap: The Math Behind This Feast Never Worked
The deeper logic behind the bond selloff is an arithmetic problem that has never had a solution.
A report released this week by technology consulting firm Gartner reveals that AI data center operators need to generate US$8.5 trillion in revenue by the end of 2030 to cover capital and financing costs and achieve healthy returns; but actual consumer and enterprise AI service spending expected to flow to data center providers is "no more than US$3.6 trillion"—a gap of US$4.9 trillion, with more than US$4 trillion in asset value potentially facing impairment risk before 2032.
Gartner further notes that the AI-optimized IaaS market is expected to grow from US$66 billion in 2027 to US$174 billion in 2030, but before existing chips are replaced by a new generation, this scale remains "insufficient to recoup investment" for most AI compute providers. More critically, because most AI-native companies adopt asset-light models, "less than half of AI revenue will flow to providers with significant AI infrastructure."
Another trigger for the shift in market sentiment was a series of signals from OpenAI: annualized revenue was US$20 billion lower than previously implied, its IPO has been delayed to next year, and Anthropic's quarterly sales growth is also slowing. Together, these signals point to a conclusion that unsettles the market: the pace of AI monetization is far behind the pace of infrastructure investment.
D.A. Davidson analyst Gil Luria put it more bluntly: CoreWeave's borrowing cost is about 9%, while its return on investment is only about 1%. "Once the AI spending boom cools, equity value will go to zero."
Nvidia CEO Jensen Huang said on Thursday that he does not believe CoreWeave needs Nvidia's support, saying "their cost of capital is actually coming down." But he provided no specific financial data, and credit market pricing stands in sharp contrast to that optimistic remark.
Two Hidden Dangers: Circular Financing and Political Backlash
Beyond the revenue gap, two types of structural risk are compounding and amplifying the severity of this crisis.
The first is circular financing risk. A recent report by the Bank for International Settlements noted that "a significant proportion of financing and commercial relationships in the AI sector are self-referential"—customers and suppliers finance each other. This structure can amplify growth when markets have tailwinds, but once confidence reverses, it may accelerate deleveraging and trigger chain reactions.
The second is political resistance from local communities. Backlash against data centers has spread across many parts of the United States—New York, Texas, and other states have imposed strict restrictions on new projects, leaving some projects stalled. Tom Murphy, head of US investment-grade credit at Columbia Threadneedle, said this makes "data centers that are already built and operating more valuable," but it also means permitting risk for new projects is being systematically underestimated.
The Strong Grow Stronger: Who Survives After the Shakeout
Despite the market turmoil, the endgame of this crisis may not be the end of AI infrastructure investment, but rather a brutal survival-of-the-fittest process.
Gartner predicts that 80% of neoclouds will "cease independent operations" before 2030—of every five companies renting Nvidia GPUs, four will be acquired or exit the market. At the same time, however, Lucinda Jerogin, an economist at Commonwealth Bank of Australia, noted that Australia's data center installed capacity has grown from about 6 gigawatts in February this year to 23 gigawatts, though only about a quarter of that is approved or under construction, and the industry's long-term growth logic remains intact.
Jun Bei Liu believes that established companies with mature financial backing and long-term agreements, such as Goodman and AirTrunk, as well as construction and engineering firms holding firm contracts, remain "true gems" worth watching. "They will continue to benefit from this generational growth, but for the 'get rich quick' players, this path will become increasingly difficult."
This is the core logic of this "Darwinian moment": in the face of a US$4.9 trillion revenue-and-expense chasm, the Firmus-style valuation game has reached its end. Only operators with real revenue, scale, and execution capability can survive this elimination round—while players that rely on circular financing and prop up valuations with stories are being liquidated one by one by the market.
Comments